Hi,
the new pipe implementation (which is much faster as well as more
kernel memory friendly) has been stabilized considerably recently.
It's likely it would become the default pipe implementation
instead current socketpair()-based ones soon.

If possible, please try out the new pipe implementation on your
system. I'd be particularily interested on how well it's working
on systems with big page size, like 16KB or more (IIRC such are on
walnut and arm26). Please send-pr any encountered problems - your
feedback is extremely important to make sure the code would
not become default before it's time (and that remaining bugs
would be found and fixed).

You'd need most up-to-date 1.5Y kernel, with sys/kern/sys_pipe.c
revision 1.14 or later. If you haven't already, switch to 
new pipe implementation by including this kernel option
to your kernel config:

options NEW_PIPE

then configure and recompile the kernel.

Hopefully you'd note performance improvement :) If you want
to benchmark, I found e.g. hbench-OS's bw_pipe useful. Note
you'd need at least hbench-OS-1.0nb1 from pkgsrc - there was
a bug in way it handled pipe descriptors, which could
make hbench deadlock under some circumstances. hbench-OS-1.0nb1
includes a fix for this.

Regards,

Jaromir
-- 
Jaromir Dolecek <jdolecek@NetBSD.org>      http://www.ics.muni.cz/~dolecek/
NetBSD - just plain best OS! -=*=- Got spare MCA cards or docs? Hand me them!